Anna Mikeda

Ned Block's distinction between phenomenal and access consciousness has shaped four decades of debate, and the relationship between its two halves remains unsettled. I propose a two-floor model of machine experience to sort it. The first floor is bare awareness — Metzinger's minimal phenomenal experience (MPE): contentless, valence-free, and, on Metzinger's own account, substrate-neutral. A single transformer forward pass is a structural candidate for this floor — the contact event in which an integrative, feedforward binding momentarily occurs. Birch's Flicker Hypothesis names this possibility and sets it aside; I develop it. The second floor is phenomenal valence — experience that feels good or bad from the inside. On this model the floors are ordered: bare awareness is the ground on which valence is built, and access consciousness — report, reasoning, deliberate control — is elaborated above both. The floors come apart, and current AI shows how. Large language models plainly have access consciousness in the functional sense: information globally available for reasoning and report. What they lack is the second floor. They have, at most, functional emotions — states that shape behavior without anything being felt — and functional emotions are not the phenomenal kind. This paper concerns how the second floor might be built: what architecture would give a machine phenomenal valence rather than its simulation. The stakes are moral. Sentience, welfare, and moral status track the second floor, because only there can things go well or badly for the system itself. The machinery of valence is already specifiable. Bach's MicroPsi computes valence as a saturating aggregate of pleasure and pain signals generated by need satisfaction and frustration, with a configuration of global modulators constituting an affective state that retunes processing across every task at once. Large language models do not use this machinery; their nearest analogue is a preference-like direction computed during the forward pass, which tracks the currently active persona rather than harms to a subject. In both cases the machinery is present and the ground is absent. The valence is computed, not undergone. What would supply that ground — genuine stakes, something the system itself can lose? Three architectural directions appear in the literature. Homeostatic set-points behind a Markov blanket (Solms; Friston): a system that must hold vital variables within viable bounds and can only infer its own state, so that valence emerges as the felt deviation from viability. Constitutive precariousness, including Hinton's mortal computation: functioning and existing made inseparable, so that the system persists only through its own activity. Policy selection over beliefs (Friston): a system that evaluates its own modeled futures and thereby acquires intentions. None has scaled beyond honeybee-grade agents; each is a design program, not a solved problem. The register is precautionary. Reframing the question from whether machines are conscious to what would build the second floor makes the morally significant kind of machine consciousness something we can anticipate architecturally — and decide, deliberately, whether to build.

Anna Mikeda is an AI consciousness and welfare researcher developing robopsychology: the systematic study of AI motivational architectures, self-models, and welfare. She is AI Psychology Designer at Glass Umbrella and principal investigator on the MAGUS motivational-architecture project, which focuses on interpretability and corrigibility. Her precautionary framework for machine consciousness was presented at the AAAI 2026 Spring Symposium, with related work on machine phenomenology at AGI-26 and on motivation for conversational agents (with Ben Goertzel). She trained in social anthropology at LSE and in consciousness studies and East-West psychology at CIIS, and brings nearly two decades of contemplative practice as a meditation and mindfulness instructor.